xref: /openbmc/linux/fs/aio.c (revision ff4a7481c3898ffc3cc271d6aca431d190c37247)
1 /*
2  *	An async IO implementation for Linux
3  *	Written by Benjamin LaHaise <bcrl@kvack.org>
4  *
5  *	Implements an efficient asynchronous io interface.
6  *
7  *	Copyright 2000, 2001, 2002 Red Hat, Inc.  All Rights Reserved.
8  *	Copyright 2018 Christoph Hellwig.
9  *
10  *	See ../COPYING for licensing terms.
11  */
12 #define pr_fmt(fmt) "%s: " fmt, __func__
13 
14 #include <linux/kernel.h>
15 #include <linux/init.h>
16 #include <linux/errno.h>
17 #include <linux/time.h>
18 #include <linux/aio_abi.h>
19 #include <linux/export.h>
20 #include <linux/syscalls.h>
21 #include <linux/backing-dev.h>
22 #include <linux/refcount.h>
23 #include <linux/uio.h>
24 
25 #include <linux/sched/signal.h>
26 #include <linux/fs.h>
27 #include <linux/file.h>
28 #include <linux/mm.h>
29 #include <linux/mman.h>
30 #include <linux/mmu_context.h>
31 #include <linux/percpu.h>
32 #include <linux/slab.h>
33 #include <linux/timer.h>
34 #include <linux/aio.h>
35 #include <linux/highmem.h>
36 #include <linux/workqueue.h>
37 #include <linux/security.h>
38 #include <linux/eventfd.h>
39 #include <linux/blkdev.h>
40 #include <linux/compat.h>
41 #include <linux/migrate.h>
42 #include <linux/ramfs.h>
43 #include <linux/percpu-refcount.h>
44 #include <linux/mount.h>
45 
46 #include <asm/kmap_types.h>
47 #include <linux/uaccess.h>
48 #include <linux/nospec.h>
49 
50 #include "internal.h"
51 
52 #define KIOCB_KEY		0
53 
54 #define AIO_RING_MAGIC			0xa10a10a1
55 #define AIO_RING_COMPAT_FEATURES	1
56 #define AIO_RING_INCOMPAT_FEATURES	0
57 struct aio_ring {
58 	unsigned	id;	/* kernel internal index number */
59 	unsigned	nr;	/* number of io_events */
60 	unsigned	head;	/* Written to by userland or under ring_lock
61 				 * mutex by aio_read_events_ring(). */
62 	unsigned	tail;
63 
64 	unsigned	magic;
65 	unsigned	compat_features;
66 	unsigned	incompat_features;
67 	unsigned	header_length;	/* size of aio_ring */
68 
69 
70 	struct io_event		io_events[0];
71 }; /* 128 bytes + ring size */
72 
73 #define AIO_RING_PAGES	8
74 
75 struct kioctx_table {
76 	struct rcu_head		rcu;
77 	unsigned		nr;
78 	struct kioctx __rcu	*table[];
79 };
80 
81 struct kioctx_cpu {
82 	unsigned		reqs_available;
83 };
84 
85 struct ctx_rq_wait {
86 	struct completion comp;
87 	atomic_t count;
88 };
89 
90 struct kioctx {
91 	struct percpu_ref	users;
92 	atomic_t		dead;
93 
94 	struct percpu_ref	reqs;
95 
96 	unsigned long		user_id;
97 
98 	struct __percpu kioctx_cpu *cpu;
99 
100 	/*
101 	 * For percpu reqs_available, number of slots we move to/from global
102 	 * counter at a time:
103 	 */
104 	unsigned		req_batch;
105 	/*
106 	 * This is what userspace passed to io_setup(), it's not used for
107 	 * anything but counting against the global max_reqs quota.
108 	 *
109 	 * The real limit is nr_events - 1, which will be larger (see
110 	 * aio_setup_ring())
111 	 */
112 	unsigned		max_reqs;
113 
114 	/* Size of ringbuffer, in units of struct io_event */
115 	unsigned		nr_events;
116 
117 	unsigned long		mmap_base;
118 	unsigned long		mmap_size;
119 
120 	struct page		**ring_pages;
121 	long			nr_pages;
122 
123 	struct rcu_work		free_rwork;	/* see free_ioctx() */
124 
125 	/*
126 	 * signals when all in-flight requests are done
127 	 */
128 	struct ctx_rq_wait	*rq_wait;
129 
130 	struct {
131 		/*
132 		 * This counts the number of available slots in the ringbuffer,
133 		 * so we avoid overflowing it: it's decremented (if positive)
134 		 * when allocating a kiocb and incremented when the resulting
135 		 * io_event is pulled off the ringbuffer.
136 		 *
137 		 * We batch accesses to it with a percpu version.
138 		 */
139 		atomic_t	reqs_available;
140 	} ____cacheline_aligned_in_smp;
141 
142 	struct {
143 		spinlock_t	ctx_lock;
144 		struct list_head active_reqs;	/* used for cancellation */
145 	} ____cacheline_aligned_in_smp;
146 
147 	struct {
148 		struct mutex	ring_lock;
149 		wait_queue_head_t wait;
150 	} ____cacheline_aligned_in_smp;
151 
152 	struct {
153 		unsigned	tail;
154 		unsigned	completed_events;
155 		spinlock_t	completion_lock;
156 	} ____cacheline_aligned_in_smp;
157 
158 	struct page		*internal_pages[AIO_RING_PAGES];
159 	struct file		*aio_ring_file;
160 
161 	unsigned		id;
162 };
163 
164 struct fsync_iocb {
165 	struct work_struct	work;
166 	struct file		*file;
167 	bool			datasync;
168 };
169 
170 struct poll_iocb {
171 	struct file		*file;
172 	struct wait_queue_head	*head;
173 	__poll_t		events;
174 	bool			woken;
175 	bool			cancelled;
176 	struct wait_queue_entry	wait;
177 	struct work_struct	work;
178 };
179 
180 struct aio_kiocb {
181 	union {
182 		struct kiocb		rw;
183 		struct fsync_iocb	fsync;
184 		struct poll_iocb	poll;
185 	};
186 
187 	struct kioctx		*ki_ctx;
188 	kiocb_cancel_fn		*ki_cancel;
189 
190 	struct iocb __user	*ki_user_iocb;	/* user's aiocb */
191 	__u64			ki_user_data;	/* user's data for completion */
192 
193 	struct list_head	ki_list;	/* the aio core uses this
194 						 * for cancellation */
195 	refcount_t		ki_refcnt;
196 
197 	/*
198 	 * If the aio_resfd field of the userspace iocb is not zero,
199 	 * this is the underlying eventfd context to deliver events to.
200 	 */
201 	struct eventfd_ctx	*ki_eventfd;
202 };
203 
204 /*------ sysctl variables----*/
205 static DEFINE_SPINLOCK(aio_nr_lock);
206 unsigned long aio_nr;		/* current system wide number of aio requests */
207 unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio requests */
208 /*----end sysctl variables---*/
209 
210 static struct kmem_cache	*kiocb_cachep;
211 static struct kmem_cache	*kioctx_cachep;
212 
213 static struct vfsmount *aio_mnt;
214 
215 static const struct file_operations aio_ring_fops;
216 static const struct address_space_operations aio_ctx_aops;
217 
218 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
219 {
220 	struct file *file;
221 	struct inode *inode = alloc_anon_inode(aio_mnt->mnt_sb);
222 	if (IS_ERR(inode))
223 		return ERR_CAST(inode);
224 
225 	inode->i_mapping->a_ops = &aio_ctx_aops;
226 	inode->i_mapping->private_data = ctx;
227 	inode->i_size = PAGE_SIZE * nr_pages;
228 
229 	file = alloc_file_pseudo(inode, aio_mnt, "[aio]",
230 				O_RDWR, &aio_ring_fops);
231 	if (IS_ERR(file))
232 		iput(inode);
233 	return file;
234 }
235 
236 static struct dentry *aio_mount(struct file_system_type *fs_type,
237 				int flags, const char *dev_name, void *data)
238 {
239 	struct dentry *root = mount_pseudo(fs_type, "aio:", NULL, NULL,
240 					   AIO_RING_MAGIC);
241 
242 	if (!IS_ERR(root))
243 		root->d_sb->s_iflags |= SB_I_NOEXEC;
244 	return root;
245 }
246 
247 /* aio_setup
248  *	Creates the slab caches used by the aio routines, panic on
249  *	failure as this is done early during the boot sequence.
250  */
251 static int __init aio_setup(void)
252 {
253 	static struct file_system_type aio_fs = {
254 		.name		= "aio",
255 		.mount		= aio_mount,
256 		.kill_sb	= kill_anon_super,
257 	};
258 	aio_mnt = kern_mount(&aio_fs);
259 	if (IS_ERR(aio_mnt))
260 		panic("Failed to create aio fs mount.");
261 
262 	kiocb_cachep = KMEM_CACHE(aio_kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
263 	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
264 	return 0;
265 }
266 __initcall(aio_setup);
267 
268 static void put_aio_ring_file(struct kioctx *ctx)
269 {
270 	struct file *aio_ring_file = ctx->aio_ring_file;
271 	struct address_space *i_mapping;
272 
273 	if (aio_ring_file) {
274 		truncate_setsize(file_inode(aio_ring_file), 0);
275 
276 		/* Prevent further access to the kioctx from migratepages */
277 		i_mapping = aio_ring_file->f_mapping;
278 		spin_lock(&i_mapping->private_lock);
279 		i_mapping->private_data = NULL;
280 		ctx->aio_ring_file = NULL;
281 		spin_unlock(&i_mapping->private_lock);
282 
283 		fput(aio_ring_file);
284 	}
285 }
286 
287 static void aio_free_ring(struct kioctx *ctx)
288 {
289 	int i;
290 
291 	/* Disconnect the kiotx from the ring file.  This prevents future
292 	 * accesses to the kioctx from page migration.
293 	 */
294 	put_aio_ring_file(ctx);
295 
296 	for (i = 0; i < ctx->nr_pages; i++) {
297 		struct page *page;
298 		pr_debug("pid(%d) [%d] page->count=%d\n", current->pid, i,
299 				page_count(ctx->ring_pages[i]));
300 		page = ctx->ring_pages[i];
301 		if (!page)
302 			continue;
303 		ctx->ring_pages[i] = NULL;
304 		put_page(page);
305 	}
306 
307 	if (ctx->ring_pages && ctx->ring_pages != ctx->internal_pages) {
308 		kfree(ctx->ring_pages);
309 		ctx->ring_pages = NULL;
310 	}
311 }
312 
313 static int aio_ring_mremap(struct vm_area_struct *vma)
314 {
315 	struct file *file = vma->vm_file;
316 	struct mm_struct *mm = vma->vm_mm;
317 	struct kioctx_table *table;
318 	int i, res = -EINVAL;
319 
320 	spin_lock(&mm->ioctx_lock);
321 	rcu_read_lock();
322 	table = rcu_dereference(mm->ioctx_table);
323 	for (i = 0; i < table->nr; i++) {
324 		struct kioctx *ctx;
325 
326 		ctx = rcu_dereference(table->table[i]);
327 		if (ctx && ctx->aio_ring_file == file) {
328 			if (!atomic_read(&ctx->dead)) {
329 				ctx->user_id = ctx->mmap_base = vma->vm_start;
330 				res = 0;
331 			}
332 			break;
333 		}
334 	}
335 
336 	rcu_read_unlock();
337 	spin_unlock(&mm->ioctx_lock);
338 	return res;
339 }
340 
341 static const struct vm_operations_struct aio_ring_vm_ops = {
342 	.mremap		= aio_ring_mremap,
343 #if IS_ENABLED(CONFIG_MMU)
344 	.fault		= filemap_fault,
345 	.map_pages	= filemap_map_pages,
346 	.page_mkwrite	= filemap_page_mkwrite,
347 #endif
348 };
349 
350 static int aio_ring_mmap(struct file *file, struct vm_area_struct *vma)
351 {
352 	vma->vm_flags |= VM_DONTEXPAND;
353 	vma->vm_ops = &aio_ring_vm_ops;
354 	return 0;
355 }
356 
357 static const struct file_operations aio_ring_fops = {
358 	.mmap = aio_ring_mmap,
359 };
360 
361 #if IS_ENABLED(CONFIG_MIGRATION)
362 static int aio_migratepage(struct address_space *mapping, struct page *new,
363 			struct page *old, enum migrate_mode mode)
364 {
365 	struct kioctx *ctx;
366 	unsigned long flags;
367 	pgoff_t idx;
368 	int rc;
369 
370 	/*
371 	 * We cannot support the _NO_COPY case here, because copy needs to
372 	 * happen under the ctx->completion_lock. That does not work with the
373 	 * migration workflow of MIGRATE_SYNC_NO_COPY.
374 	 */
375 	if (mode == MIGRATE_SYNC_NO_COPY)
376 		return -EINVAL;
377 
378 	rc = 0;
379 
380 	/* mapping->private_lock here protects against the kioctx teardown.  */
381 	spin_lock(&mapping->private_lock);
382 	ctx = mapping->private_data;
383 	if (!ctx) {
384 		rc = -EINVAL;
385 		goto out;
386 	}
387 
388 	/* The ring_lock mutex.  The prevents aio_read_events() from writing
389 	 * to the ring's head, and prevents page migration from mucking in
390 	 * a partially initialized kiotx.
391 	 */
392 	if (!mutex_trylock(&ctx->ring_lock)) {
393 		rc = -EAGAIN;
394 		goto out;
395 	}
396 
397 	idx = old->index;
398 	if (idx < (pgoff_t)ctx->nr_pages) {
399 		/* Make sure the old page hasn't already been changed */
400 		if (ctx->ring_pages[idx] != old)
401 			rc = -EAGAIN;
402 	} else
403 		rc = -EINVAL;
404 
405 	if (rc != 0)
406 		goto out_unlock;
407 
408 	/* Writeback must be complete */
409 	BUG_ON(PageWriteback(old));
410 	get_page(new);
411 
412 	rc = migrate_page_move_mapping(mapping, new, old, NULL, mode, 1);
413 	if (rc != MIGRATEPAGE_SUCCESS) {
414 		put_page(new);
415 		goto out_unlock;
416 	}
417 
418 	/* Take completion_lock to prevent other writes to the ring buffer
419 	 * while the old page is copied to the new.  This prevents new
420 	 * events from being lost.
421 	 */
422 	spin_lock_irqsave(&ctx->completion_lock, flags);
423 	migrate_page_copy(new, old);
424 	BUG_ON(ctx->ring_pages[idx] != old);
425 	ctx->ring_pages[idx] = new;
426 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
427 
428 	/* The old page is no longer accessible. */
429 	put_page(old);
430 
431 out_unlock:
432 	mutex_unlock(&ctx->ring_lock);
433 out:
434 	spin_unlock(&mapping->private_lock);
435 	return rc;
436 }
437 #endif
438 
439 static const struct address_space_operations aio_ctx_aops = {
440 	.set_page_dirty = __set_page_dirty_no_writeback,
441 #if IS_ENABLED(CONFIG_MIGRATION)
442 	.migratepage	= aio_migratepage,
443 #endif
444 };
445 
446 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
447 {
448 	struct aio_ring *ring;
449 	struct mm_struct *mm = current->mm;
450 	unsigned long size, unused;
451 	int nr_pages;
452 	int i;
453 	struct file *file;
454 
455 	/* Compensate for the ring buffer's head/tail overlap entry */
456 	nr_events += 2;	/* 1 is required, 2 for good luck */
457 
458 	size = sizeof(struct aio_ring);
459 	size += sizeof(struct io_event) * nr_events;
460 
461 	nr_pages = PFN_UP(size);
462 	if (nr_pages < 0)
463 		return -EINVAL;
464 
465 	file = aio_private_file(ctx, nr_pages);
466 	if (IS_ERR(file)) {
467 		ctx->aio_ring_file = NULL;
468 		return -ENOMEM;
469 	}
470 
471 	ctx->aio_ring_file = file;
472 	nr_events = (PAGE_SIZE * nr_pages - sizeof(struct aio_ring))
473 			/ sizeof(struct io_event);
474 
475 	ctx->ring_pages = ctx->internal_pages;
476 	if (nr_pages > AIO_RING_PAGES) {
477 		ctx->ring_pages = kcalloc(nr_pages, sizeof(struct page *),
478 					  GFP_KERNEL);
479 		if (!ctx->ring_pages) {
480 			put_aio_ring_file(ctx);
481 			return -ENOMEM;
482 		}
483 	}
484 
485 	for (i = 0; i < nr_pages; i++) {
486 		struct page *page;
487 		page = find_or_create_page(file->f_mapping,
488 					   i, GFP_HIGHUSER | __GFP_ZERO);
489 		if (!page)
490 			break;
491 		pr_debug("pid(%d) page[%d]->count=%d\n",
492 			 current->pid, i, page_count(page));
493 		SetPageUptodate(page);
494 		unlock_page(page);
495 
496 		ctx->ring_pages[i] = page;
497 	}
498 	ctx->nr_pages = i;
499 
500 	if (unlikely(i != nr_pages)) {
501 		aio_free_ring(ctx);
502 		return -ENOMEM;
503 	}
504 
505 	ctx->mmap_size = nr_pages * PAGE_SIZE;
506 	pr_debug("attempting mmap of %lu bytes\n", ctx->mmap_size);
507 
508 	if (down_write_killable(&mm->mmap_sem)) {
509 		ctx->mmap_size = 0;
510 		aio_free_ring(ctx);
511 		return -EINTR;
512 	}
513 
514 	ctx->mmap_base = do_mmap_pgoff(ctx->aio_ring_file, 0, ctx->mmap_size,
515 				       PROT_READ | PROT_WRITE,
516 				       MAP_SHARED, 0, &unused, NULL);
517 	up_write(&mm->mmap_sem);
518 	if (IS_ERR((void *)ctx->mmap_base)) {
519 		ctx->mmap_size = 0;
520 		aio_free_ring(ctx);
521 		return -ENOMEM;
522 	}
523 
524 	pr_debug("mmap address: 0x%08lx\n", ctx->mmap_base);
525 
526 	ctx->user_id = ctx->mmap_base;
527 	ctx->nr_events = nr_events; /* trusted copy */
528 
529 	ring = kmap_atomic(ctx->ring_pages[0]);
530 	ring->nr = nr_events;	/* user copy */
531 	ring->id = ~0U;
532 	ring->head = ring->tail = 0;
533 	ring->magic = AIO_RING_MAGIC;
534 	ring->compat_features = AIO_RING_COMPAT_FEATURES;
535 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
536 	ring->header_length = sizeof(struct aio_ring);
537 	kunmap_atomic(ring);
538 	flush_dcache_page(ctx->ring_pages[0]);
539 
540 	return 0;
541 }
542 
543 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
544 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
545 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
546 
547 void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
548 {
549 	struct aio_kiocb *req = container_of(iocb, struct aio_kiocb, rw);
550 	struct kioctx *ctx = req->ki_ctx;
551 	unsigned long flags;
552 
553 	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
554 		return;
555 
556 	spin_lock_irqsave(&ctx->ctx_lock, flags);
557 	list_add_tail(&req->ki_list, &ctx->active_reqs);
558 	req->ki_cancel = cancel;
559 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
560 }
561 EXPORT_SYMBOL(kiocb_set_cancel_fn);
562 
563 /*
564  * free_ioctx() should be RCU delayed to synchronize against the RCU
565  * protected lookup_ioctx() and also needs process context to call
566  * aio_free_ring().  Use rcu_work.
567  */
568 static void free_ioctx(struct work_struct *work)
569 {
570 	struct kioctx *ctx = container_of(to_rcu_work(work), struct kioctx,
571 					  free_rwork);
572 	pr_debug("freeing %p\n", ctx);
573 
574 	aio_free_ring(ctx);
575 	free_percpu(ctx->cpu);
576 	percpu_ref_exit(&ctx->reqs);
577 	percpu_ref_exit(&ctx->users);
578 	kmem_cache_free(kioctx_cachep, ctx);
579 }
580 
581 static void free_ioctx_reqs(struct percpu_ref *ref)
582 {
583 	struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
584 
585 	/* At this point we know that there are no any in-flight requests */
586 	if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
587 		complete(&ctx->rq_wait->comp);
588 
589 	/* Synchronize against RCU protected table->table[] dereferences */
590 	INIT_RCU_WORK(&ctx->free_rwork, free_ioctx);
591 	queue_rcu_work(system_wq, &ctx->free_rwork);
592 }
593 
594 /*
595  * When this function runs, the kioctx has been removed from the "hash table"
596  * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
597  * now it's safe to cancel any that need to be.
598  */
599 static void free_ioctx_users(struct percpu_ref *ref)
600 {
601 	struct kioctx *ctx = container_of(ref, struct kioctx, users);
602 	struct aio_kiocb *req;
603 
604 	spin_lock_irq(&ctx->ctx_lock);
605 
606 	while (!list_empty(&ctx->active_reqs)) {
607 		req = list_first_entry(&ctx->active_reqs,
608 				       struct aio_kiocb, ki_list);
609 		req->ki_cancel(&req->rw);
610 		list_del_init(&req->ki_list);
611 	}
612 
613 	spin_unlock_irq(&ctx->ctx_lock);
614 
615 	percpu_ref_kill(&ctx->reqs);
616 	percpu_ref_put(&ctx->reqs);
617 }
618 
619 static int ioctx_add_table(struct kioctx *ctx, struct mm_struct *mm)
620 {
621 	unsigned i, new_nr;
622 	struct kioctx_table *table, *old;
623 	struct aio_ring *ring;
624 
625 	spin_lock(&mm->ioctx_lock);
626 	table = rcu_dereference_raw(mm->ioctx_table);
627 
628 	while (1) {
629 		if (table)
630 			for (i = 0; i < table->nr; i++)
631 				if (!rcu_access_pointer(table->table[i])) {
632 					ctx->id = i;
633 					rcu_assign_pointer(table->table[i], ctx);
634 					spin_unlock(&mm->ioctx_lock);
635 
636 					/* While kioctx setup is in progress,
637 					 * we are protected from page migration
638 					 * changes ring_pages by ->ring_lock.
639 					 */
640 					ring = kmap_atomic(ctx->ring_pages[0]);
641 					ring->id = ctx->id;
642 					kunmap_atomic(ring);
643 					return 0;
644 				}
645 
646 		new_nr = (table ? table->nr : 1) * 4;
647 		spin_unlock(&mm->ioctx_lock);
648 
649 		table = kzalloc(sizeof(*table) + sizeof(struct kioctx *) *
650 				new_nr, GFP_KERNEL);
651 		if (!table)
652 			return -ENOMEM;
653 
654 		table->nr = new_nr;
655 
656 		spin_lock(&mm->ioctx_lock);
657 		old = rcu_dereference_raw(mm->ioctx_table);
658 
659 		if (!old) {
660 			rcu_assign_pointer(mm->ioctx_table, table);
661 		} else if (table->nr > old->nr) {
662 			memcpy(table->table, old->table,
663 			       old->nr * sizeof(struct kioctx *));
664 
665 			rcu_assign_pointer(mm->ioctx_table, table);
666 			kfree_rcu(old, rcu);
667 		} else {
668 			kfree(table);
669 			table = old;
670 		}
671 	}
672 }
673 
674 static void aio_nr_sub(unsigned nr)
675 {
676 	spin_lock(&aio_nr_lock);
677 	if (WARN_ON(aio_nr - nr > aio_nr))
678 		aio_nr = 0;
679 	else
680 		aio_nr -= nr;
681 	spin_unlock(&aio_nr_lock);
682 }
683 
684 /* ioctx_alloc
685  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
686  */
687 static struct kioctx *ioctx_alloc(unsigned nr_events)
688 {
689 	struct mm_struct *mm = current->mm;
690 	struct kioctx *ctx;
691 	int err = -ENOMEM;
692 
693 	/*
694 	 * Store the original nr_events -- what userspace passed to io_setup(),
695 	 * for counting against the global limit -- before it changes.
696 	 */
697 	unsigned int max_reqs = nr_events;
698 
699 	/*
700 	 * We keep track of the number of available ringbuffer slots, to prevent
701 	 * overflow (reqs_available), and we also use percpu counters for this.
702 	 *
703 	 * So since up to half the slots might be on other cpu's percpu counters
704 	 * and unavailable, double nr_events so userspace sees what they
705 	 * expected: additionally, we move req_batch slots to/from percpu
706 	 * counters at a time, so make sure that isn't 0:
707 	 */
708 	nr_events = max(nr_events, num_possible_cpus() * 4);
709 	nr_events *= 2;
710 
711 	/* Prevent overflows */
712 	if (nr_events > (0x10000000U / sizeof(struct io_event))) {
713 		pr_debug("ENOMEM: nr_events too high\n");
714 		return ERR_PTR(-EINVAL);
715 	}
716 
717 	if (!nr_events || (unsigned long)max_reqs > aio_max_nr)
718 		return ERR_PTR(-EAGAIN);
719 
720 	ctx = kmem_cache_zalloc(kioctx_cachep, GFP_KERNEL);
721 	if (!ctx)
722 		return ERR_PTR(-ENOMEM);
723 
724 	ctx->max_reqs = max_reqs;
725 
726 	spin_lock_init(&ctx->ctx_lock);
727 	spin_lock_init(&ctx->completion_lock);
728 	mutex_init(&ctx->ring_lock);
729 	/* Protect against page migration throughout kiotx setup by keeping
730 	 * the ring_lock mutex held until setup is complete. */
731 	mutex_lock(&ctx->ring_lock);
732 	init_waitqueue_head(&ctx->wait);
733 
734 	INIT_LIST_HEAD(&ctx->active_reqs);
735 
736 	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
737 		goto err;
738 
739 	if (percpu_ref_init(&ctx->reqs, free_ioctx_reqs, 0, GFP_KERNEL))
740 		goto err;
741 
742 	ctx->cpu = alloc_percpu(struct kioctx_cpu);
743 	if (!ctx->cpu)
744 		goto err;
745 
746 	err = aio_setup_ring(ctx, nr_events);
747 	if (err < 0)
748 		goto err;
749 
750 	atomic_set(&ctx->reqs_available, ctx->nr_events - 1);
751 	ctx->req_batch = (ctx->nr_events - 1) / (num_possible_cpus() * 4);
752 	if (ctx->req_batch < 1)
753 		ctx->req_batch = 1;
754 
755 	/* limit the number of system wide aios */
756 	spin_lock(&aio_nr_lock);
757 	if (aio_nr + ctx->max_reqs > aio_max_nr ||
758 	    aio_nr + ctx->max_reqs < aio_nr) {
759 		spin_unlock(&aio_nr_lock);
760 		err = -EAGAIN;
761 		goto err_ctx;
762 	}
763 	aio_nr += ctx->max_reqs;
764 	spin_unlock(&aio_nr_lock);
765 
766 	percpu_ref_get(&ctx->users);	/* io_setup() will drop this ref */
767 	percpu_ref_get(&ctx->reqs);	/* free_ioctx_users() will drop this */
768 
769 	err = ioctx_add_table(ctx, mm);
770 	if (err)
771 		goto err_cleanup;
772 
773 	/* Release the ring_lock mutex now that all setup is complete. */
774 	mutex_unlock(&ctx->ring_lock);
775 
776 	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
777 		 ctx, ctx->user_id, mm, ctx->nr_events);
778 	return ctx;
779 
780 err_cleanup:
781 	aio_nr_sub(ctx->max_reqs);
782 err_ctx:
783 	atomic_set(&ctx->dead, 1);
784 	if (ctx->mmap_size)
785 		vm_munmap(ctx->mmap_base, ctx->mmap_size);
786 	aio_free_ring(ctx);
787 err:
788 	mutex_unlock(&ctx->ring_lock);
789 	free_percpu(ctx->cpu);
790 	percpu_ref_exit(&ctx->reqs);
791 	percpu_ref_exit(&ctx->users);
792 	kmem_cache_free(kioctx_cachep, ctx);
793 	pr_debug("error allocating ioctx %d\n", err);
794 	return ERR_PTR(err);
795 }
796 
797 /* kill_ioctx
798  *	Cancels all outstanding aio requests on an aio context.  Used
799  *	when the processes owning a context have all exited to encourage
800  *	the rapid destruction of the kioctx.
801  */
802 static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
803 		      struct ctx_rq_wait *wait)
804 {
805 	struct kioctx_table *table;
806 
807 	spin_lock(&mm->ioctx_lock);
808 	if (atomic_xchg(&ctx->dead, 1)) {
809 		spin_unlock(&mm->ioctx_lock);
810 		return -EINVAL;
811 	}
812 
813 	table = rcu_dereference_raw(mm->ioctx_table);
814 	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
815 	RCU_INIT_POINTER(table->table[ctx->id], NULL);
816 	spin_unlock(&mm->ioctx_lock);
817 
818 	/* free_ioctx_reqs() will do the necessary RCU synchronization */
819 	wake_up_all(&ctx->wait);
820 
821 	/*
822 	 * It'd be more correct to do this in free_ioctx(), after all
823 	 * the outstanding kiocbs have finished - but by then io_destroy
824 	 * has already returned, so io_setup() could potentially return
825 	 * -EAGAIN with no ioctxs actually in use (as far as userspace
826 	 *  could tell).
827 	 */
828 	aio_nr_sub(ctx->max_reqs);
829 
830 	if (ctx->mmap_size)
831 		vm_munmap(ctx->mmap_base, ctx->mmap_size);
832 
833 	ctx->rq_wait = wait;
834 	percpu_ref_kill(&ctx->users);
835 	return 0;
836 }
837 
838 /*
839  * exit_aio: called when the last user of mm goes away.  At this point, there is
840  * no way for any new requests to be submited or any of the io_* syscalls to be
841  * called on the context.
842  *
843  * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
844  * them.
845  */
846 void exit_aio(struct mm_struct *mm)
847 {
848 	struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
849 	struct ctx_rq_wait wait;
850 	int i, skipped;
851 
852 	if (!table)
853 		return;
854 
855 	atomic_set(&wait.count, table->nr);
856 	init_completion(&wait.comp);
857 
858 	skipped = 0;
859 	for (i = 0; i < table->nr; ++i) {
860 		struct kioctx *ctx =
861 			rcu_dereference_protected(table->table[i], true);
862 
863 		if (!ctx) {
864 			skipped++;
865 			continue;
866 		}
867 
868 		/*
869 		 * We don't need to bother with munmap() here - exit_mmap(mm)
870 		 * is coming and it'll unmap everything. And we simply can't,
871 		 * this is not necessarily our ->mm.
872 		 * Since kill_ioctx() uses non-zero ->mmap_size as indicator
873 		 * that it needs to unmap the area, just set it to 0.
874 		 */
875 		ctx->mmap_size = 0;
876 		kill_ioctx(mm, ctx, &wait);
877 	}
878 
879 	if (!atomic_sub_and_test(skipped, &wait.count)) {
880 		/* Wait until all IO for the context are done. */
881 		wait_for_completion(&wait.comp);
882 	}
883 
884 	RCU_INIT_POINTER(mm->ioctx_table, NULL);
885 	kfree(table);
886 }
887 
888 static void put_reqs_available(struct kioctx *ctx, unsigned nr)
889 {
890 	struct kioctx_cpu *kcpu;
891 	unsigned long flags;
892 
893 	local_irq_save(flags);
894 	kcpu = this_cpu_ptr(ctx->cpu);
895 	kcpu->reqs_available += nr;
896 
897 	while (kcpu->reqs_available >= ctx->req_batch * 2) {
898 		kcpu->reqs_available -= ctx->req_batch;
899 		atomic_add(ctx->req_batch, &ctx->reqs_available);
900 	}
901 
902 	local_irq_restore(flags);
903 }
904 
905 static bool get_reqs_available(struct kioctx *ctx)
906 {
907 	struct kioctx_cpu *kcpu;
908 	bool ret = false;
909 	unsigned long flags;
910 
911 	local_irq_save(flags);
912 	kcpu = this_cpu_ptr(ctx->cpu);
913 	if (!kcpu->reqs_available) {
914 		int old, avail = atomic_read(&ctx->reqs_available);
915 
916 		do {
917 			if (avail < ctx->req_batch)
918 				goto out;
919 
920 			old = avail;
921 			avail = atomic_cmpxchg(&ctx->reqs_available,
922 					       avail, avail - ctx->req_batch);
923 		} while (avail != old);
924 
925 		kcpu->reqs_available += ctx->req_batch;
926 	}
927 
928 	ret = true;
929 	kcpu->reqs_available--;
930 out:
931 	local_irq_restore(flags);
932 	return ret;
933 }
934 
935 /* refill_reqs_available
936  *	Updates the reqs_available reference counts used for tracking the
937  *	number of free slots in the completion ring.  This can be called
938  *	from aio_complete() (to optimistically update reqs_available) or
939  *	from aio_get_req() (the we're out of events case).  It must be
940  *	called holding ctx->completion_lock.
941  */
942 static void refill_reqs_available(struct kioctx *ctx, unsigned head,
943                                   unsigned tail)
944 {
945 	unsigned events_in_ring, completed;
946 
947 	/* Clamp head since userland can write to it. */
948 	head %= ctx->nr_events;
949 	if (head <= tail)
950 		events_in_ring = tail - head;
951 	else
952 		events_in_ring = ctx->nr_events - (head - tail);
953 
954 	completed = ctx->completed_events;
955 	if (events_in_ring < completed)
956 		completed -= events_in_ring;
957 	else
958 		completed = 0;
959 
960 	if (!completed)
961 		return;
962 
963 	ctx->completed_events -= completed;
964 	put_reqs_available(ctx, completed);
965 }
966 
967 /* user_refill_reqs_available
968  *	Called to refill reqs_available when aio_get_req() encounters an
969  *	out of space in the completion ring.
970  */
971 static void user_refill_reqs_available(struct kioctx *ctx)
972 {
973 	spin_lock_irq(&ctx->completion_lock);
974 	if (ctx->completed_events) {
975 		struct aio_ring *ring;
976 		unsigned head;
977 
978 		/* Access of ring->head may race with aio_read_events_ring()
979 		 * here, but that's okay since whether we read the old version
980 		 * or the new version, and either will be valid.  The important
981 		 * part is that head cannot pass tail since we prevent
982 		 * aio_complete() from updating tail by holding
983 		 * ctx->completion_lock.  Even if head is invalid, the check
984 		 * against ctx->completed_events below will make sure we do the
985 		 * safe/right thing.
986 		 */
987 		ring = kmap_atomic(ctx->ring_pages[0]);
988 		head = ring->head;
989 		kunmap_atomic(ring);
990 
991 		refill_reqs_available(ctx, head, ctx->tail);
992 	}
993 
994 	spin_unlock_irq(&ctx->completion_lock);
995 }
996 
997 /* aio_get_req
998  *	Allocate a slot for an aio request.
999  * Returns NULL if no requests are free.
1000  */
1001 static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
1002 {
1003 	struct aio_kiocb *req;
1004 
1005 	if (!get_reqs_available(ctx)) {
1006 		user_refill_reqs_available(ctx);
1007 		if (!get_reqs_available(ctx))
1008 			return NULL;
1009 	}
1010 
1011 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
1012 	if (unlikely(!req))
1013 		goto out_put;
1014 
1015 	percpu_ref_get(&ctx->reqs);
1016 	INIT_LIST_HEAD(&req->ki_list);
1017 	refcount_set(&req->ki_refcnt, 0);
1018 	req->ki_ctx = ctx;
1019 	return req;
1020 out_put:
1021 	put_reqs_available(ctx, 1);
1022 	return NULL;
1023 }
1024 
1025 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
1026 {
1027 	struct aio_ring __user *ring  = (void __user *)ctx_id;
1028 	struct mm_struct *mm = current->mm;
1029 	struct kioctx *ctx, *ret = NULL;
1030 	struct kioctx_table *table;
1031 	unsigned id;
1032 
1033 	if (get_user(id, &ring->id))
1034 		return NULL;
1035 
1036 	rcu_read_lock();
1037 	table = rcu_dereference(mm->ioctx_table);
1038 
1039 	if (!table || id >= table->nr)
1040 		goto out;
1041 
1042 	id = array_index_nospec(id, table->nr);
1043 	ctx = rcu_dereference(table->table[id]);
1044 	if (ctx && ctx->user_id == ctx_id) {
1045 		if (percpu_ref_tryget_live(&ctx->users))
1046 			ret = ctx;
1047 	}
1048 out:
1049 	rcu_read_unlock();
1050 	return ret;
1051 }
1052 
1053 static inline void iocb_put(struct aio_kiocb *iocb)
1054 {
1055 	if (refcount_read(&iocb->ki_refcnt) == 0 ||
1056 	    refcount_dec_and_test(&iocb->ki_refcnt)) {
1057 		percpu_ref_put(&iocb->ki_ctx->reqs);
1058 		kmem_cache_free(kiocb_cachep, iocb);
1059 	}
1060 }
1061 
1062 /* aio_complete
1063  *	Called when the io request on the given iocb is complete.
1064  */
1065 static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
1066 {
1067 	struct kioctx	*ctx = iocb->ki_ctx;
1068 	struct aio_ring	*ring;
1069 	struct io_event	*ev_page, *event;
1070 	unsigned tail, pos, head;
1071 	unsigned long	flags;
1072 
1073 	/*
1074 	 * Add a completion event to the ring buffer. Must be done holding
1075 	 * ctx->completion_lock to prevent other code from messing with the tail
1076 	 * pointer since we might be called from irq context.
1077 	 */
1078 	spin_lock_irqsave(&ctx->completion_lock, flags);
1079 
1080 	tail = ctx->tail;
1081 	pos = tail + AIO_EVENTS_OFFSET;
1082 
1083 	if (++tail >= ctx->nr_events)
1084 		tail = 0;
1085 
1086 	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1087 	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
1088 
1089 	event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
1090 	event->data = iocb->ki_user_data;
1091 	event->res = res;
1092 	event->res2 = res2;
1093 
1094 	kunmap_atomic(ev_page);
1095 	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
1096 
1097 	pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
1098 		 ctx, tail, iocb, iocb->ki_user_iocb, iocb->ki_user_data,
1099 		 res, res2);
1100 
1101 	/* after flagging the request as done, we
1102 	 * must never even look at it again
1103 	 */
1104 	smp_wmb();	/* make event visible before updating tail */
1105 
1106 	ctx->tail = tail;
1107 
1108 	ring = kmap_atomic(ctx->ring_pages[0]);
1109 	head = ring->head;
1110 	ring->tail = tail;
1111 	kunmap_atomic(ring);
1112 	flush_dcache_page(ctx->ring_pages[0]);
1113 
1114 	ctx->completed_events++;
1115 	if (ctx->completed_events > 1)
1116 		refill_reqs_available(ctx, head, tail);
1117 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
1118 
1119 	pr_debug("added to ring %p at [%u]\n", iocb, tail);
1120 
1121 	/*
1122 	 * Check if the user asked us to deliver the result through an
1123 	 * eventfd. The eventfd_signal() function is safe to be called
1124 	 * from IRQ context.
1125 	 */
1126 	if (iocb->ki_eventfd) {
1127 		eventfd_signal(iocb->ki_eventfd, 1);
1128 		eventfd_ctx_put(iocb->ki_eventfd);
1129 	}
1130 
1131 	/*
1132 	 * We have to order our ring_info tail store above and test
1133 	 * of the wait list below outside the wait lock.  This is
1134 	 * like in wake_up_bit() where clearing a bit has to be
1135 	 * ordered with the unlocked test.
1136 	 */
1137 	smp_mb();
1138 
1139 	if (waitqueue_active(&ctx->wait))
1140 		wake_up(&ctx->wait);
1141 	iocb_put(iocb);
1142 }
1143 
1144 /* aio_read_events_ring
1145  *	Pull an event off of the ioctx's event ring.  Returns the number of
1146  *	events fetched
1147  */
1148 static long aio_read_events_ring(struct kioctx *ctx,
1149 				 struct io_event __user *event, long nr)
1150 {
1151 	struct aio_ring *ring;
1152 	unsigned head, tail, pos;
1153 	long ret = 0;
1154 	int copy_ret;
1155 
1156 	/*
1157 	 * The mutex can block and wake us up and that will cause
1158 	 * wait_event_interruptible_hrtimeout() to schedule without sleeping
1159 	 * and repeat. This should be rare enough that it doesn't cause
1160 	 * peformance issues. See the comment in read_events() for more detail.
1161 	 */
1162 	sched_annotate_sleep();
1163 	mutex_lock(&ctx->ring_lock);
1164 
1165 	/* Access to ->ring_pages here is protected by ctx->ring_lock. */
1166 	ring = kmap_atomic(ctx->ring_pages[0]);
1167 	head = ring->head;
1168 	tail = ring->tail;
1169 	kunmap_atomic(ring);
1170 
1171 	/*
1172 	 * Ensure that once we've read the current tail pointer, that
1173 	 * we also see the events that were stored up to the tail.
1174 	 */
1175 	smp_rmb();
1176 
1177 	pr_debug("h%u t%u m%u\n", head, tail, ctx->nr_events);
1178 
1179 	if (head == tail)
1180 		goto out;
1181 
1182 	head %= ctx->nr_events;
1183 	tail %= ctx->nr_events;
1184 
1185 	while (ret < nr) {
1186 		long avail;
1187 		struct io_event *ev;
1188 		struct page *page;
1189 
1190 		avail = (head <= tail ?  tail : ctx->nr_events) - head;
1191 		if (head == tail)
1192 			break;
1193 
1194 		pos = head + AIO_EVENTS_OFFSET;
1195 		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
1196 		pos %= AIO_EVENTS_PER_PAGE;
1197 
1198 		avail = min(avail, nr - ret);
1199 		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE - pos);
1200 
1201 		ev = kmap(page);
1202 		copy_ret = copy_to_user(event + ret, ev + pos,
1203 					sizeof(*ev) * avail);
1204 		kunmap(page);
1205 
1206 		if (unlikely(copy_ret)) {
1207 			ret = -EFAULT;
1208 			goto out;
1209 		}
1210 
1211 		ret += avail;
1212 		head += avail;
1213 		head %= ctx->nr_events;
1214 	}
1215 
1216 	ring = kmap_atomic(ctx->ring_pages[0]);
1217 	ring->head = head;
1218 	kunmap_atomic(ring);
1219 	flush_dcache_page(ctx->ring_pages[0]);
1220 
1221 	pr_debug("%li  h%u t%u\n", ret, head, tail);
1222 out:
1223 	mutex_unlock(&ctx->ring_lock);
1224 
1225 	return ret;
1226 }
1227 
1228 static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
1229 			    struct io_event __user *event, long *i)
1230 {
1231 	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
1232 
1233 	if (ret > 0)
1234 		*i += ret;
1235 
1236 	if (unlikely(atomic_read(&ctx->dead)))
1237 		ret = -EINVAL;
1238 
1239 	if (!*i)
1240 		*i = ret;
1241 
1242 	return ret < 0 || *i >= min_nr;
1243 }
1244 
1245 static long read_events(struct kioctx *ctx, long min_nr, long nr,
1246 			struct io_event __user *event,
1247 			ktime_t until)
1248 {
1249 	long ret = 0;
1250 
1251 	/*
1252 	 * Note that aio_read_events() is being called as the conditional - i.e.
1253 	 * we're calling it after prepare_to_wait() has set task state to
1254 	 * TASK_INTERRUPTIBLE.
1255 	 *
1256 	 * But aio_read_events() can block, and if it blocks it's going to flip
1257 	 * the task state back to TASK_RUNNING.
1258 	 *
1259 	 * This should be ok, provided it doesn't flip the state back to
1260 	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
1261 	 * will only happen if the mutex_lock() call blocks, and we then find
1262 	 * the ringbuffer empty. So in practice we should be ok, but it's
1263 	 * something to be aware of when touching this code.
1264 	 */
1265 	if (until == 0)
1266 		aio_read_events(ctx, min_nr, nr, event, &ret);
1267 	else
1268 		wait_event_interruptible_hrtimeout(ctx->wait,
1269 				aio_read_events(ctx, min_nr, nr, event, &ret),
1270 				until);
1271 	return ret;
1272 }
1273 
1274 /* sys_io_setup:
1275  *	Create an aio_context capable of receiving at least nr_events.
1276  *	ctxp must not point to an aio_context that already exists, and
1277  *	must be initialized to 0 prior to the call.  On successful
1278  *	creation of the aio_context, *ctxp is filled in with the resulting
1279  *	handle.  May fail with -EINVAL if *ctxp is not initialized,
1280  *	if the specified nr_events exceeds internal limits.  May fail
1281  *	with -EAGAIN if the specified nr_events exceeds the user's limit
1282  *	of available events.  May fail with -ENOMEM if insufficient kernel
1283  *	resources are available.  May fail with -EFAULT if an invalid
1284  *	pointer is passed for ctxp.  Will fail with -ENOSYS if not
1285  *	implemented.
1286  */
1287 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
1288 {
1289 	struct kioctx *ioctx = NULL;
1290 	unsigned long ctx;
1291 	long ret;
1292 
1293 	ret = get_user(ctx, ctxp);
1294 	if (unlikely(ret))
1295 		goto out;
1296 
1297 	ret = -EINVAL;
1298 	if (unlikely(ctx || nr_events == 0)) {
1299 		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1300 		         ctx, nr_events);
1301 		goto out;
1302 	}
1303 
1304 	ioctx = ioctx_alloc(nr_events);
1305 	ret = PTR_ERR(ioctx);
1306 	if (!IS_ERR(ioctx)) {
1307 		ret = put_user(ioctx->user_id, ctxp);
1308 		if (ret)
1309 			kill_ioctx(current->mm, ioctx, NULL);
1310 		percpu_ref_put(&ioctx->users);
1311 	}
1312 
1313 out:
1314 	return ret;
1315 }
1316 
1317 #ifdef CONFIG_COMPAT
1318 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
1319 {
1320 	struct kioctx *ioctx = NULL;
1321 	unsigned long ctx;
1322 	long ret;
1323 
1324 	ret = get_user(ctx, ctx32p);
1325 	if (unlikely(ret))
1326 		goto out;
1327 
1328 	ret = -EINVAL;
1329 	if (unlikely(ctx || nr_events == 0)) {
1330 		pr_debug("EINVAL: ctx %lu nr_events %u\n",
1331 		         ctx, nr_events);
1332 		goto out;
1333 	}
1334 
1335 	ioctx = ioctx_alloc(nr_events);
1336 	ret = PTR_ERR(ioctx);
1337 	if (!IS_ERR(ioctx)) {
1338 		/* truncating is ok because it's a user address */
1339 		ret = put_user((u32)ioctx->user_id, ctx32p);
1340 		if (ret)
1341 			kill_ioctx(current->mm, ioctx, NULL);
1342 		percpu_ref_put(&ioctx->users);
1343 	}
1344 
1345 out:
1346 	return ret;
1347 }
1348 #endif
1349 
1350 /* sys_io_destroy:
1351  *	Destroy the aio_context specified.  May cancel any outstanding
1352  *	AIOs and block on completion.  Will fail with -ENOSYS if not
1353  *	implemented.  May fail with -EINVAL if the context pointed to
1354  *	is invalid.
1355  */
1356 SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
1357 {
1358 	struct kioctx *ioctx = lookup_ioctx(ctx);
1359 	if (likely(NULL != ioctx)) {
1360 		struct ctx_rq_wait wait;
1361 		int ret;
1362 
1363 		init_completion(&wait.comp);
1364 		atomic_set(&wait.count, 1);
1365 
1366 		/* Pass requests_done to kill_ioctx() where it can be set
1367 		 * in a thread-safe way. If we try to set it here then we have
1368 		 * a race condition if two io_destroy() called simultaneously.
1369 		 */
1370 		ret = kill_ioctx(current->mm, ioctx, &wait);
1371 		percpu_ref_put(&ioctx->users);
1372 
1373 		/* Wait until all IO for the context are done. Otherwise kernel
1374 		 * keep using user-space buffers even if user thinks the context
1375 		 * is destroyed.
1376 		 */
1377 		if (!ret)
1378 			wait_for_completion(&wait.comp);
1379 
1380 		return ret;
1381 	}
1382 	pr_debug("EINVAL: invalid context id\n");
1383 	return -EINVAL;
1384 }
1385 
1386 static void aio_remove_iocb(struct aio_kiocb *iocb)
1387 {
1388 	struct kioctx *ctx = iocb->ki_ctx;
1389 	unsigned long flags;
1390 
1391 	spin_lock_irqsave(&ctx->ctx_lock, flags);
1392 	list_del(&iocb->ki_list);
1393 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
1394 }
1395 
1396 static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
1397 {
1398 	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
1399 
1400 	if (!list_empty_careful(&iocb->ki_list))
1401 		aio_remove_iocb(iocb);
1402 
1403 	if (kiocb->ki_flags & IOCB_WRITE) {
1404 		struct inode *inode = file_inode(kiocb->ki_filp);
1405 
1406 		/*
1407 		 * Tell lockdep we inherited freeze protection from submission
1408 		 * thread.
1409 		 */
1410 		if (S_ISREG(inode->i_mode))
1411 			__sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
1412 		file_end_write(kiocb->ki_filp);
1413 	}
1414 
1415 	fput(kiocb->ki_filp);
1416 	aio_complete(iocb, res, res2);
1417 }
1418 
1419 static int aio_prep_rw(struct kiocb *req, struct iocb *iocb)
1420 {
1421 	int ret;
1422 
1423 	req->ki_filp = fget(iocb->aio_fildes);
1424 	if (unlikely(!req->ki_filp))
1425 		return -EBADF;
1426 	req->ki_complete = aio_complete_rw;
1427 	req->ki_pos = iocb->aio_offset;
1428 	req->ki_flags = iocb_flags(req->ki_filp);
1429 	if (iocb->aio_flags & IOCB_FLAG_RESFD)
1430 		req->ki_flags |= IOCB_EVENTFD;
1431 	req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
1432 	if (iocb->aio_flags & IOCB_FLAG_IOPRIO) {
1433 		/*
1434 		 * If the IOCB_FLAG_IOPRIO flag of aio_flags is set, then
1435 		 * aio_reqprio is interpreted as an I/O scheduling
1436 		 * class and priority.
1437 		 */
1438 		ret = ioprio_check_cap(iocb->aio_reqprio);
1439 		if (ret) {
1440 			pr_debug("aio ioprio check cap error: %d\n", ret);
1441 			fput(req->ki_filp);
1442 			return ret;
1443 		}
1444 
1445 		req->ki_ioprio = iocb->aio_reqprio;
1446 	} else
1447 		req->ki_ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_NONE, 0);
1448 
1449 	ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
1450 	if (unlikely(ret))
1451 		fput(req->ki_filp);
1452 	return ret;
1453 }
1454 
1455 static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
1456 		bool vectored, bool compat, struct iov_iter *iter)
1457 {
1458 	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
1459 	size_t len = iocb->aio_nbytes;
1460 
1461 	if (!vectored) {
1462 		ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
1463 		*iovec = NULL;
1464 		return ret;
1465 	}
1466 #ifdef CONFIG_COMPAT
1467 	if (compat)
1468 		return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
1469 				iter);
1470 #endif
1471 	return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
1472 }
1473 
1474 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
1475 {
1476 	switch (ret) {
1477 	case -EIOCBQUEUED:
1478 		break;
1479 	case -ERESTARTSYS:
1480 	case -ERESTARTNOINTR:
1481 	case -ERESTARTNOHAND:
1482 	case -ERESTART_RESTARTBLOCK:
1483 		/*
1484 		 * There's no easy way to restart the syscall since other AIO's
1485 		 * may be already running. Just fail this IO with EINTR.
1486 		 */
1487 		ret = -EINTR;
1488 		/*FALLTHRU*/
1489 	default:
1490 		aio_complete_rw(req, ret, 0);
1491 	}
1492 }
1493 
1494 static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
1495 		bool compat)
1496 {
1497 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1498 	struct iov_iter iter;
1499 	struct file *file;
1500 	ssize_t ret;
1501 
1502 	ret = aio_prep_rw(req, iocb);
1503 	if (ret)
1504 		return ret;
1505 	file = req->ki_filp;
1506 
1507 	ret = -EBADF;
1508 	if (unlikely(!(file->f_mode & FMODE_READ)))
1509 		goto out_fput;
1510 	ret = -EINVAL;
1511 	if (unlikely(!file->f_op->read_iter))
1512 		goto out_fput;
1513 
1514 	ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
1515 	if (ret)
1516 		goto out_fput;
1517 	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
1518 	if (!ret)
1519 		aio_rw_done(req, call_read_iter(file, req, &iter));
1520 	kfree(iovec);
1521 out_fput:
1522 	if (unlikely(ret))
1523 		fput(file);
1524 	return ret;
1525 }
1526 
1527 static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
1528 		bool compat)
1529 {
1530 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
1531 	struct iov_iter iter;
1532 	struct file *file;
1533 	ssize_t ret;
1534 
1535 	ret = aio_prep_rw(req, iocb);
1536 	if (ret)
1537 		return ret;
1538 	file = req->ki_filp;
1539 
1540 	ret = -EBADF;
1541 	if (unlikely(!(file->f_mode & FMODE_WRITE)))
1542 		goto out_fput;
1543 	ret = -EINVAL;
1544 	if (unlikely(!file->f_op->write_iter))
1545 		goto out_fput;
1546 
1547 	ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
1548 	if (ret)
1549 		goto out_fput;
1550 	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
1551 	if (!ret) {
1552 		/*
1553 		 * Open-code file_start_write here to grab freeze protection,
1554 		 * which will be released by another thread in
1555 		 * aio_complete_rw().  Fool lockdep by telling it the lock got
1556 		 * released so that it doesn't complain about the held lock when
1557 		 * we return to userspace.
1558 		 */
1559 		if (S_ISREG(file_inode(file)->i_mode)) {
1560 			__sb_start_write(file_inode(file)->i_sb, SB_FREEZE_WRITE, true);
1561 			__sb_writers_release(file_inode(file)->i_sb, SB_FREEZE_WRITE);
1562 		}
1563 		req->ki_flags |= IOCB_WRITE;
1564 		aio_rw_done(req, call_write_iter(file, req, &iter));
1565 	}
1566 	kfree(iovec);
1567 out_fput:
1568 	if (unlikely(ret))
1569 		fput(file);
1570 	return ret;
1571 }
1572 
1573 static void aio_fsync_work(struct work_struct *work)
1574 {
1575 	struct fsync_iocb *req = container_of(work, struct fsync_iocb, work);
1576 	int ret;
1577 
1578 	ret = vfs_fsync(req->file, req->datasync);
1579 	fput(req->file);
1580 	aio_complete(container_of(req, struct aio_kiocb, fsync), ret, 0);
1581 }
1582 
1583 static int aio_fsync(struct fsync_iocb *req, struct iocb *iocb, bool datasync)
1584 {
1585 	if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
1586 			iocb->aio_rw_flags))
1587 		return -EINVAL;
1588 
1589 	req->file = fget(iocb->aio_fildes);
1590 	if (unlikely(!req->file))
1591 		return -EBADF;
1592 	if (unlikely(!req->file->f_op->fsync)) {
1593 		fput(req->file);
1594 		return -EINVAL;
1595 	}
1596 
1597 	req->datasync = datasync;
1598 	INIT_WORK(&req->work, aio_fsync_work);
1599 	schedule_work(&req->work);
1600 	return 0;
1601 }
1602 
1603 static inline void aio_poll_complete(struct aio_kiocb *iocb, __poll_t mask)
1604 {
1605 	struct file *file = iocb->poll.file;
1606 
1607 	aio_complete(iocb, mangle_poll(mask), 0);
1608 	fput(file);
1609 }
1610 
1611 static void aio_poll_complete_work(struct work_struct *work)
1612 {
1613 	struct poll_iocb *req = container_of(work, struct poll_iocb, work);
1614 	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1615 	struct poll_table_struct pt = { ._key = req->events };
1616 	struct kioctx *ctx = iocb->ki_ctx;
1617 	__poll_t mask = 0;
1618 
1619 	if (!READ_ONCE(req->cancelled))
1620 		mask = vfs_poll(req->file, &pt) & req->events;
1621 
1622 	/*
1623 	 * Note that ->ki_cancel callers also delete iocb from active_reqs after
1624 	 * calling ->ki_cancel.  We need the ctx_lock roundtrip here to
1625 	 * synchronize with them.  In the cancellation case the list_del_init
1626 	 * itself is not actually needed, but harmless so we keep it in to
1627 	 * avoid further branches in the fast path.
1628 	 */
1629 	spin_lock_irq(&ctx->ctx_lock);
1630 	if (!mask && !READ_ONCE(req->cancelled)) {
1631 		add_wait_queue(req->head, &req->wait);
1632 		spin_unlock_irq(&ctx->ctx_lock);
1633 		return;
1634 	}
1635 	list_del_init(&iocb->ki_list);
1636 	spin_unlock_irq(&ctx->ctx_lock);
1637 
1638 	aio_poll_complete(iocb, mask);
1639 }
1640 
1641 /* assumes we are called with irqs disabled */
1642 static int aio_poll_cancel(struct kiocb *iocb)
1643 {
1644 	struct aio_kiocb *aiocb = container_of(iocb, struct aio_kiocb, rw);
1645 	struct poll_iocb *req = &aiocb->poll;
1646 
1647 	spin_lock(&req->head->lock);
1648 	WRITE_ONCE(req->cancelled, true);
1649 	if (!list_empty(&req->wait.entry)) {
1650 		list_del_init(&req->wait.entry);
1651 		schedule_work(&aiocb->poll.work);
1652 	}
1653 	spin_unlock(&req->head->lock);
1654 
1655 	return 0;
1656 }
1657 
1658 static int aio_poll_wake(struct wait_queue_entry *wait, unsigned mode, int sync,
1659 		void *key)
1660 {
1661 	struct poll_iocb *req = container_of(wait, struct poll_iocb, wait);
1662 	struct aio_kiocb *iocb = container_of(req, struct aio_kiocb, poll);
1663 	__poll_t mask = key_to_poll(key);
1664 
1665 	req->woken = true;
1666 
1667 	/* for instances that support it check for an event match first: */
1668 	if (mask) {
1669 		if (!(mask & req->events))
1670 			return 0;
1671 
1672 		/* try to complete the iocb inline if we can: */
1673 		if (spin_trylock(&iocb->ki_ctx->ctx_lock)) {
1674 			list_del(&iocb->ki_list);
1675 			spin_unlock(&iocb->ki_ctx->ctx_lock);
1676 
1677 			list_del_init(&req->wait.entry);
1678 			aio_poll_complete(iocb, mask);
1679 			return 1;
1680 		}
1681 	}
1682 
1683 	list_del_init(&req->wait.entry);
1684 	schedule_work(&req->work);
1685 	return 1;
1686 }
1687 
1688 struct aio_poll_table {
1689 	struct poll_table_struct	pt;
1690 	struct aio_kiocb		*iocb;
1691 	int				error;
1692 };
1693 
1694 static void
1695 aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
1696 		struct poll_table_struct *p)
1697 {
1698 	struct aio_poll_table *pt = container_of(p, struct aio_poll_table, pt);
1699 
1700 	/* multiple wait queues per file are not supported */
1701 	if (unlikely(pt->iocb->poll.head)) {
1702 		pt->error = -EINVAL;
1703 		return;
1704 	}
1705 
1706 	pt->error = 0;
1707 	pt->iocb->poll.head = head;
1708 	add_wait_queue(head, &pt->iocb->poll.wait);
1709 }
1710 
1711 static ssize_t aio_poll(struct aio_kiocb *aiocb, struct iocb *iocb)
1712 {
1713 	struct kioctx *ctx = aiocb->ki_ctx;
1714 	struct poll_iocb *req = &aiocb->poll;
1715 	struct aio_poll_table apt;
1716 	__poll_t mask;
1717 
1718 	/* reject any unknown events outside the normal event mask. */
1719 	if ((u16)iocb->aio_buf != iocb->aio_buf)
1720 		return -EINVAL;
1721 	/* reject fields that are not defined for poll */
1722 	if (iocb->aio_offset || iocb->aio_nbytes || iocb->aio_rw_flags)
1723 		return -EINVAL;
1724 
1725 	INIT_WORK(&req->work, aio_poll_complete_work);
1726 	req->events = demangle_poll(iocb->aio_buf) | EPOLLERR | EPOLLHUP;
1727 	req->file = fget(iocb->aio_fildes);
1728 	if (unlikely(!req->file))
1729 		return -EBADF;
1730 
1731 	apt.pt._qproc = aio_poll_queue_proc;
1732 	apt.pt._key = req->events;
1733 	apt.iocb = aiocb;
1734 	apt.error = -EINVAL; /* same as no support for IOCB_CMD_POLL */
1735 
1736 	/* initialized the list so that we can do list_empty checks */
1737 	INIT_LIST_HEAD(&req->wait.entry);
1738 	init_waitqueue_func_entry(&req->wait, aio_poll_wake);
1739 
1740 	/* one for removal from waitqueue, one for this function */
1741 	refcount_set(&aiocb->ki_refcnt, 2);
1742 
1743 	mask = vfs_poll(req->file, &apt.pt) & req->events;
1744 	if (unlikely(!req->head)) {
1745 		/* we did not manage to set up a waitqueue, done */
1746 		goto out;
1747 	}
1748 
1749 	spin_lock_irq(&ctx->ctx_lock);
1750 	spin_lock(&req->head->lock);
1751 	if (req->woken) {
1752 		/* wake_up context handles the rest */
1753 		mask = 0;
1754 		apt.error = 0;
1755 	} else if (mask || apt.error) {
1756 		/* if we get an error or a mask we are done */
1757 		WARN_ON_ONCE(list_empty(&req->wait.entry));
1758 		list_del_init(&req->wait.entry);
1759 	} else {
1760 		/* actually waiting for an event */
1761 		list_add_tail(&aiocb->ki_list, &ctx->active_reqs);
1762 		aiocb->ki_cancel = aio_poll_cancel;
1763 	}
1764 	spin_unlock(&req->head->lock);
1765 	spin_unlock_irq(&ctx->ctx_lock);
1766 
1767 out:
1768 	if (unlikely(apt.error)) {
1769 		fput(req->file);
1770 		return apt.error;
1771 	}
1772 
1773 	if (mask)
1774 		aio_poll_complete(aiocb, mask);
1775 	iocb_put(aiocb);
1776 	return 0;
1777 }
1778 
1779 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
1780 			 bool compat)
1781 {
1782 	struct aio_kiocb *req;
1783 	struct iocb iocb;
1784 	ssize_t ret;
1785 
1786 	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
1787 		return -EFAULT;
1788 
1789 	/* enforce forwards compatibility on users */
1790 	if (unlikely(iocb.aio_reserved2)) {
1791 		pr_debug("EINVAL: reserve field set\n");
1792 		return -EINVAL;
1793 	}
1794 
1795 	/* prevent overflows */
1796 	if (unlikely(
1797 	    (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
1798 	    (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
1799 	    ((ssize_t)iocb.aio_nbytes < 0)
1800 	   )) {
1801 		pr_debug("EINVAL: overflow check\n");
1802 		return -EINVAL;
1803 	}
1804 
1805 	req = aio_get_req(ctx);
1806 	if (unlikely(!req))
1807 		return -EAGAIN;
1808 
1809 	if (iocb.aio_flags & IOCB_FLAG_RESFD) {
1810 		/*
1811 		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
1812 		 * instance of the file* now. The file descriptor must be
1813 		 * an eventfd() fd, and will be signaled for each completed
1814 		 * event using the eventfd_signal() function.
1815 		 */
1816 		req->ki_eventfd = eventfd_ctx_fdget((int) iocb.aio_resfd);
1817 		if (IS_ERR(req->ki_eventfd)) {
1818 			ret = PTR_ERR(req->ki_eventfd);
1819 			req->ki_eventfd = NULL;
1820 			goto out_put_req;
1821 		}
1822 	}
1823 
1824 	ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
1825 	if (unlikely(ret)) {
1826 		pr_debug("EFAULT: aio_key\n");
1827 		goto out_put_req;
1828 	}
1829 
1830 	req->ki_user_iocb = user_iocb;
1831 	req->ki_user_data = iocb.aio_data;
1832 
1833 	switch (iocb.aio_lio_opcode) {
1834 	case IOCB_CMD_PREAD:
1835 		ret = aio_read(&req->rw, &iocb, false, compat);
1836 		break;
1837 	case IOCB_CMD_PWRITE:
1838 		ret = aio_write(&req->rw, &iocb, false, compat);
1839 		break;
1840 	case IOCB_CMD_PREADV:
1841 		ret = aio_read(&req->rw, &iocb, true, compat);
1842 		break;
1843 	case IOCB_CMD_PWRITEV:
1844 		ret = aio_write(&req->rw, &iocb, true, compat);
1845 		break;
1846 	case IOCB_CMD_FSYNC:
1847 		ret = aio_fsync(&req->fsync, &iocb, false);
1848 		break;
1849 	case IOCB_CMD_FDSYNC:
1850 		ret = aio_fsync(&req->fsync, &iocb, true);
1851 		break;
1852 	case IOCB_CMD_POLL:
1853 		ret = aio_poll(req, &iocb);
1854 		break;
1855 	default:
1856 		pr_debug("invalid aio operation %d\n", iocb.aio_lio_opcode);
1857 		ret = -EINVAL;
1858 		break;
1859 	}
1860 
1861 	/*
1862 	 * If ret is 0, we'd either done aio_complete() ourselves or have
1863 	 * arranged for that to be done asynchronously.  Anything non-zero
1864 	 * means that we need to destroy req ourselves.
1865 	 */
1866 	if (ret)
1867 		goto out_put_req;
1868 	return 0;
1869 out_put_req:
1870 	put_reqs_available(ctx, 1);
1871 	percpu_ref_put(&ctx->reqs);
1872 	if (req->ki_eventfd)
1873 		eventfd_ctx_put(req->ki_eventfd);
1874 	kmem_cache_free(kiocb_cachep, req);
1875 	return ret;
1876 }
1877 
1878 /* sys_io_submit:
1879  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
1880  *	the number of iocbs queued.  May return -EINVAL if the aio_context
1881  *	specified by ctx_id is invalid, if nr is < 0, if the iocb at
1882  *	*iocbpp[0] is not properly initialized, if the operation specified
1883  *	is invalid for the file descriptor in the iocb.  May fail with
1884  *	-EFAULT if any of the data structures point to invalid data.  May
1885  *	fail with -EBADF if the file descriptor specified in the first
1886  *	iocb is invalid.  May fail with -EAGAIN if insufficient resources
1887  *	are available to queue any iocbs.  Will return 0 if nr is 0.  Will
1888  *	fail with -ENOSYS if not implemented.
1889  */
1890 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
1891 		struct iocb __user * __user *, iocbpp)
1892 {
1893 	struct kioctx *ctx;
1894 	long ret = 0;
1895 	int i = 0;
1896 	struct blk_plug plug;
1897 
1898 	if (unlikely(nr < 0))
1899 		return -EINVAL;
1900 
1901 	ctx = lookup_ioctx(ctx_id);
1902 	if (unlikely(!ctx)) {
1903 		pr_debug("EINVAL: invalid context id\n");
1904 		return -EINVAL;
1905 	}
1906 
1907 	if (nr > ctx->nr_events)
1908 		nr = ctx->nr_events;
1909 
1910 	blk_start_plug(&plug);
1911 	for (i = 0; i < nr; i++) {
1912 		struct iocb __user *user_iocb;
1913 
1914 		if (unlikely(get_user(user_iocb, iocbpp + i))) {
1915 			ret = -EFAULT;
1916 			break;
1917 		}
1918 
1919 		ret = io_submit_one(ctx, user_iocb, false);
1920 		if (ret)
1921 			break;
1922 	}
1923 	blk_finish_plug(&plug);
1924 
1925 	percpu_ref_put(&ctx->users);
1926 	return i ? i : ret;
1927 }
1928 
1929 #ifdef CONFIG_COMPAT
1930 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
1931 		       int, nr, compat_uptr_t __user *, iocbpp)
1932 {
1933 	struct kioctx *ctx;
1934 	long ret = 0;
1935 	int i = 0;
1936 	struct blk_plug plug;
1937 
1938 	if (unlikely(nr < 0))
1939 		return -EINVAL;
1940 
1941 	ctx = lookup_ioctx(ctx_id);
1942 	if (unlikely(!ctx)) {
1943 		pr_debug("EINVAL: invalid context id\n");
1944 		return -EINVAL;
1945 	}
1946 
1947 	if (nr > ctx->nr_events)
1948 		nr = ctx->nr_events;
1949 
1950 	blk_start_plug(&plug);
1951 	for (i = 0; i < nr; i++) {
1952 		compat_uptr_t user_iocb;
1953 
1954 		if (unlikely(get_user(user_iocb, iocbpp + i))) {
1955 			ret = -EFAULT;
1956 			break;
1957 		}
1958 
1959 		ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
1960 		if (ret)
1961 			break;
1962 	}
1963 	blk_finish_plug(&plug);
1964 
1965 	percpu_ref_put(&ctx->users);
1966 	return i ? i : ret;
1967 }
1968 #endif
1969 
1970 /* lookup_kiocb
1971  *	Finds a given iocb for cancellation.
1972  */
1973 static struct aio_kiocb *
1974 lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb)
1975 {
1976 	struct aio_kiocb *kiocb;
1977 
1978 	assert_spin_locked(&ctx->ctx_lock);
1979 
1980 	/* TODO: use a hash or array, this sucks. */
1981 	list_for_each_entry(kiocb, &ctx->active_reqs, ki_list) {
1982 		if (kiocb->ki_user_iocb == iocb)
1983 			return kiocb;
1984 	}
1985 	return NULL;
1986 }
1987 
1988 /* sys_io_cancel:
1989  *	Attempts to cancel an iocb previously passed to io_submit.  If
1990  *	the operation is successfully cancelled, the resulting event is
1991  *	copied into the memory pointed to by result without being placed
1992  *	into the completion queue and 0 is returned.  May fail with
1993  *	-EFAULT if any of the data structures pointed to are invalid.
1994  *	May fail with -EINVAL if aio_context specified by ctx_id is
1995  *	invalid.  May fail with -EAGAIN if the iocb specified was not
1996  *	cancelled.  Will fail with -ENOSYS if not implemented.
1997  */
1998 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
1999 		struct io_event __user *, result)
2000 {
2001 	struct kioctx *ctx;
2002 	struct aio_kiocb *kiocb;
2003 	int ret = -EINVAL;
2004 	u32 key;
2005 
2006 	if (unlikely(get_user(key, &iocb->aio_key)))
2007 		return -EFAULT;
2008 	if (unlikely(key != KIOCB_KEY))
2009 		return -EINVAL;
2010 
2011 	ctx = lookup_ioctx(ctx_id);
2012 	if (unlikely(!ctx))
2013 		return -EINVAL;
2014 
2015 	spin_lock_irq(&ctx->ctx_lock);
2016 	kiocb = lookup_kiocb(ctx, iocb);
2017 	if (kiocb) {
2018 		ret = kiocb->ki_cancel(&kiocb->rw);
2019 		list_del_init(&kiocb->ki_list);
2020 	}
2021 	spin_unlock_irq(&ctx->ctx_lock);
2022 
2023 	if (!ret) {
2024 		/*
2025 		 * The result argument is no longer used - the io_event is
2026 		 * always delivered via the ring buffer. -EINPROGRESS indicates
2027 		 * cancellation is progress:
2028 		 */
2029 		ret = -EINPROGRESS;
2030 	}
2031 
2032 	percpu_ref_put(&ctx->users);
2033 
2034 	return ret;
2035 }
2036 
2037 static long do_io_getevents(aio_context_t ctx_id,
2038 		long min_nr,
2039 		long nr,
2040 		struct io_event __user *events,
2041 		struct timespec64 *ts)
2042 {
2043 	ktime_t until = ts ? timespec64_to_ktime(*ts) : KTIME_MAX;
2044 	struct kioctx *ioctx = lookup_ioctx(ctx_id);
2045 	long ret = -EINVAL;
2046 
2047 	if (likely(ioctx)) {
2048 		if (likely(min_nr <= nr && min_nr >= 0))
2049 			ret = read_events(ioctx, min_nr, nr, events, until);
2050 		percpu_ref_put(&ioctx->users);
2051 	}
2052 
2053 	return ret;
2054 }
2055 
2056 /* io_getevents:
2057  *	Attempts to read at least min_nr events and up to nr events from
2058  *	the completion queue for the aio_context specified by ctx_id. If
2059  *	it succeeds, the number of read events is returned. May fail with
2060  *	-EINVAL if ctx_id is invalid, if min_nr is out of range, if nr is
2061  *	out of range, if timeout is out of range.  May fail with -EFAULT
2062  *	if any of the memory specified is invalid.  May return 0 or
2063  *	< min_nr if the timeout specified by timeout has elapsed
2064  *	before sufficient events are available, where timeout == NULL
2065  *	specifies an infinite timeout. Note that the timeout pointed to by
2066  *	timeout is relative.  Will fail with -ENOSYS if not implemented.
2067  */
2068 SYSCALL_DEFINE5(io_getevents, aio_context_t, ctx_id,
2069 		long, min_nr,
2070 		long, nr,
2071 		struct io_event __user *, events,
2072 		struct timespec __user *, timeout)
2073 {
2074 	struct timespec64	ts;
2075 	int			ret;
2076 
2077 	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2078 		return -EFAULT;
2079 
2080 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2081 	if (!ret && signal_pending(current))
2082 		ret = -EINTR;
2083 	return ret;
2084 }
2085 
2086 struct __aio_sigset {
2087 	const sigset_t __user	*sigmask;
2088 	size_t		sigsetsize;
2089 };
2090 
2091 SYSCALL_DEFINE6(io_pgetevents,
2092 		aio_context_t, ctx_id,
2093 		long, min_nr,
2094 		long, nr,
2095 		struct io_event __user *, events,
2096 		struct timespec __user *, timeout,
2097 		const struct __aio_sigset __user *, usig)
2098 {
2099 	struct __aio_sigset	ksig = { NULL, };
2100 	sigset_t		ksigmask, sigsaved;
2101 	struct timespec64	ts;
2102 	int ret;
2103 
2104 	if (timeout && unlikely(get_timespec64(&ts, timeout)))
2105 		return -EFAULT;
2106 
2107 	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2108 		return -EFAULT;
2109 
2110 	if (ksig.sigmask) {
2111 		if (ksig.sigsetsize != sizeof(sigset_t))
2112 			return -EINVAL;
2113 		if (copy_from_user(&ksigmask, ksig.sigmask, sizeof(ksigmask)))
2114 			return -EFAULT;
2115 		sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
2116 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
2117 	}
2118 
2119 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &ts : NULL);
2120 	if (signal_pending(current)) {
2121 		if (ksig.sigmask) {
2122 			current->saved_sigmask = sigsaved;
2123 			set_restore_sigmask();
2124 		}
2125 
2126 		if (!ret)
2127 			ret = -ERESTARTNOHAND;
2128 	} else {
2129 		if (ksig.sigmask)
2130 			sigprocmask(SIG_SETMASK, &sigsaved, NULL);
2131 	}
2132 
2133 	return ret;
2134 }
2135 
2136 #ifdef CONFIG_COMPAT
2137 COMPAT_SYSCALL_DEFINE5(io_getevents, compat_aio_context_t, ctx_id,
2138 		       compat_long_t, min_nr,
2139 		       compat_long_t, nr,
2140 		       struct io_event __user *, events,
2141 		       struct old_timespec32 __user *, timeout)
2142 {
2143 	struct timespec64 t;
2144 	int ret;
2145 
2146 	if (timeout && get_old_timespec32(&t, timeout))
2147 		return -EFAULT;
2148 
2149 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2150 	if (!ret && signal_pending(current))
2151 		ret = -EINTR;
2152 	return ret;
2153 }
2154 
2155 
2156 struct __compat_aio_sigset {
2157 	compat_sigset_t __user	*sigmask;
2158 	compat_size_t		sigsetsize;
2159 };
2160 
2161 COMPAT_SYSCALL_DEFINE6(io_pgetevents,
2162 		compat_aio_context_t, ctx_id,
2163 		compat_long_t, min_nr,
2164 		compat_long_t, nr,
2165 		struct io_event __user *, events,
2166 		struct old_timespec32 __user *, timeout,
2167 		const struct __compat_aio_sigset __user *, usig)
2168 {
2169 	struct __compat_aio_sigset ksig = { NULL, };
2170 	sigset_t ksigmask, sigsaved;
2171 	struct timespec64 t;
2172 	int ret;
2173 
2174 	if (timeout && get_old_timespec32(&t, timeout))
2175 		return -EFAULT;
2176 
2177 	if (usig && copy_from_user(&ksig, usig, sizeof(ksig)))
2178 		return -EFAULT;
2179 
2180 	if (ksig.sigmask) {
2181 		if (ksig.sigsetsize != sizeof(compat_sigset_t))
2182 			return -EINVAL;
2183 		if (get_compat_sigset(&ksigmask, ksig.sigmask))
2184 			return -EFAULT;
2185 		sigdelsetmask(&ksigmask, sigmask(SIGKILL) | sigmask(SIGSTOP));
2186 		sigprocmask(SIG_SETMASK, &ksigmask, &sigsaved);
2187 	}
2188 
2189 	ret = do_io_getevents(ctx_id, min_nr, nr, events, timeout ? &t : NULL);
2190 	if (signal_pending(current)) {
2191 		if (ksig.sigmask) {
2192 			current->saved_sigmask = sigsaved;
2193 			set_restore_sigmask();
2194 		}
2195 		if (!ret)
2196 			ret = -ERESTARTNOHAND;
2197 	} else {
2198 		if (ksig.sigmask)
2199 			sigprocmask(SIG_SETMASK, &sigsaved, NULL);
2200 	}
2201 
2202 	return ret;
2203 }
2204 #endif
2205